Graph-based localization of impact noise in buildings

This study focuses on localizing impact noise in buildings, particularly noise that disturbs residents. Traditional data-driven approaches using supervised learning with location labeling often neglect the structural influence of buildings, which act as complex mediums for sound transmission. Previous research revealed reduced localization performance when the same source-to-receiver positions were tested across different floor sections [Appl. Sci. 11, 5399 (2021)]. To overcome this limitation, this study proposes a graph-based localization method that incorporates building structure. The building’s components are modeled as graph edges, spatial relationships as nodes, and medium properties as vertex features. The model’s learnable weights are optimized to minimize the discrepancy between observed impact noise and predicted responses at receiver nodes. The method is validated using two data splits with identical source-to-receiver positions but different floor sections. Visual analysis highlights distributional differences between splits. Transfer learning tasks demonstrate that the graph-based method outperforms baseline models in knowledge transfer. Additionally, visual demonstrations confirm its superior localization performance.